Sovereign AI & Private RAG

Air-Gapped Local RAG for Engineering Schematics & Hardware Manuals

Client identifiers anonymized under strict Non-Disclosure Agreements (NDAs). Performance metrics reflect architectural benchmark simulations.

About the Project

This system was engineered for an industrial manufacturing and defense hardware facility that produces precision mechanical components. Technicians and assembly engineers work with thousands of dense technical schematics, wiring diagrams, torque specifications, and operating procedures.

Because the facilities operate under strict military and industrial security protocols, all workstations are completely disconnected from the internet. The goal was to provide engineers with a high-accuracy technical assistant running 100% locally on on-premise hardware without external network pings or cloud dependencies.

Challenges We Faced

1. Strict Zero-Trust Physical & Network Air-Gap

The platform had to be deployed into server racks with physically severed network cards. No cloud APIs (OpenAI, Anthropic, AWS) could be queried, package managers could not download external dependencies, and all models, tokenizers, and database engines had to run entirely self-contained.

2. Complex Tabular Matrices & Engineering Schematics

Industrial manuals contain multi-column mechanical tolerances, wiring diagrams, and exploded component diagrams. Standard OCR tools completely failed on tabular matrices, reading horizontally across columns and scrambling critical measurement tolerances.

3. Strict VRAM Limits on Local Hardware

Each factory workstation rack was equipped with 4x NVIDIA RTX 6000 Ada GPUs (48GB VRAM each). Loading large unquantized foundation models caused immediate out-of-memory crashes when multiple engineers submitted concurrent complex queries.

Solution Architecture

Acadify architected a fully self-contained, air-gapped RAG appliance delivered through our enterprise AI deployment solutions practice.

We packaged quantized open-source models using 4-bit Activation-aware Weight Quantization (AWQ), allowing high-throughput local inference on existing workstation hardware without degrading technical comprehension.

  • Layout-Aware Document Ingestion: Engineered a local parsing pipeline using Microsoft Table Transformer and PyMuPDF to extract tables and technical diagrams into structured JSON objects, preserving exact row-column relationships and numerical tolerances.
  • Local Embedded Vector Store: Deployed an air-gapped Milvus instance running bge-base-en-v1.5 embeddings on dedicated GPU cores, delivering candidate retrieval in under 120ms without any network egress.
  • Quantized Local Inference: Deployed Mistral NeMo 12B and Llama 3 8B Instruct models running on vLLM with 4-bit AWQ, fitting comfortably within workstation memory bounds while supporting up to 8 concurrent engineering queries.
  • Visual Citation Overlays: Responses included bounding-box visual coordinates mapped directly onto original schematic PDF previews, allowing technicians to verify measurements against the physical technical drawing.

Why Acadify's Engineering Approach Fit the Project

Deploying generative AI in true air-gapped environments requires deep bare-metal engineering knowledge—managing CUDA memory allocations, local container orchestration, and offline model quantization. Acadify built a hardened local system that respects physical isolation boundaries while delivering instant technical precision.

System Architecture & Tech Stack
Runtime: Local Docker Engine (Air-Gapped)  |  Inference: vLLM (4-bit AWQ)  |  Layout Parsing: Table Transformer & PyMuPDF  |  Vector Engine: Milvus (Self-Hosted)  |  Models: Mistral NeMo 12B & Llama 3 8B  |  Hardware: 4x NVIDIA RTX 6000 Ada (48GB).

Related Acadify Solution Services

This case study demonstrates capabilities from Acadify Solution's AI Development and Enterprise AI Deployment practices, specifically air-gapped RAG architecture for sensitive environments. For organizations adopting AI in regulated settings, see our Enterprise AI Adoption roadmap.

Engineering-Led AI Evaluation & Enterprise Software

Ready to Architect, Evaluate, or Scale Your AI Systems?

From RAG evaluations, model training, and agent benchmarking at our dedicated practice Acadify AI to private VPC deployment and full-stack software development at Acadify Solution. 100% IP ownership, mutual NDAs, and deterministic failure analysis.

Strict Mutual NDA in 24h
100% Client IP Ownership
4h+ Daily US Overlap (PST/EST)
Zero Data Retention & Private VPC